Searching Efficient Model-Guided Deep Network for Image Denoising.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Neural Architecture Search (NAS) excels in high-level vision but struggles with efficient low-level vision solutions.
• Differential NAS for image restoration faces optimization gaps, leading to search instability.

Purpose of the Study:

• To address the challenges of efficient and stable NAS for low-level vision tasks, specifically image denoising.
• To introduce a novel approach, Model-Guided NAS (MoD-NAS), that bridges the optimization gap.

Main Methods:

• Developed MoD-NAS by integrating model-guided design with NAS, creating a new search space.
• Implemented stable and efficient differential search strategies, including reusable width search and densely connected blocks.
• Utilized gradient descent for automatic selection of layer operations, network width, and depth.

Main Results:

• MoD-NAS demonstrated stable search processes due to a smoother, model-guided search space.
• Achieved comparable or superior Peak Signal-to-Noise Ratio (PSNR) performance compared to state-of-the-art methods.
• Resulted in models with fewer parameters, lower Floating Point Operations (FLOPs), and reduced testing time.

Conclusions:

• MoD-NAS offers a stable and efficient solution for neural architecture search in image denoising.
• The proposed method achieves high performance with significant reductions in computational cost and resource requirements.
• MoD-NAS provides a promising direction for developing efficient deep learning models for low-level vision tasks.